AI Advances Poised to Revamp TB Diagnosis and Care

Researchers unveiled at the The Union World Conference on Lung Health 2025 in Copenhagen a wave of artificial-intelligence tools designed to speed up detection, refine treatment and improve targeting of care for tuberculosis (TB). The disclosures spotlight novel machine-learning methods applied to breath analysis, cough audio and vulnerability mapping—each aiming to close critical gaps in diagnosis and follow-up for this longstanding infectious disease.

One major innovation involves non-invasive monitoring of treatment progress through breath-chemical analysis. A collaborative team from the Southern University of Science and Technology and Shenzhen Third People’s Hospital collected exhaled breath samples from people undergoing TB therapy and fed the data into machine-learning models. They report that this method indicated response to treatment months earlier than standard sputum tests or chest imaging. According to pulmonologist Liang Fu of the Shenzhen hospital, “A non-invasive breath test combined with machine learning can track recovery during TB treatment, and indicate when a patient is likely doing well before the standard six months.” The potential to shorten therapy durations, improve adherence and cut costs is framed as a step toward precision TB management.

Another line of investigation focuses on mobile cough-sound analysis. An Indian startup, Salcit Technologies—leveraging the HeAR acoustic model developed by Google Research—has developed its product “Swaasa” which analyses cough recordings via smartphone. The model achieved about 72 % sensitivity and 71 % specificity in preliminary tests, indicating promise for low-cost, community-based screening in areas with limited access to traditional diagnostics. Lead researcher Sujay Kakarmath described how the acoustic biomarker approach “could make care more accessible and efficient.”

Thirdly, health-systems researchers from Wadhwani Institute for AI and the country’s central TB programme developed a vulnerability-mapping tool that uses over twenty open-source datasets including demographic, geographic and economic indicators, combined with national TB surveillance records. The AI model achieved 71 % accuracy in identifying the top 20 % of villages most likely to contain undiagnosed TB cases—facilitating more efficient, targeted active case-finding campaigns in hard-to-reach communities.

Beyond these specific studies, a broader review of AI in TB control shows the field is rapidly evolving. A scoping review of 34 studies found AI applications across screenings, diagnostics and monitoring of TB. The review points out that while most research to date has concentrated on static imaging and detection, emerging efforts are branching into dynamic monitoring and real-world deployment challenges. A separate meta-analysis of five commercial computer-aided detection tools found variability in performance across different country settings and emphasised the need for scenario-specific calibration of thresholds to deliver reliable results in high-burden contexts.

Analysts caution, however, that translation from study to scale involves key hurdles. Data-bias concerns remain acute: most algorithm training has occurred in limited settings, raising questions about generalisability to diverse populations and health systems. Regulatory approvals are patchy and few tools are yet endorsed by major regulatory bodies. A landscape assessment by the global diagnostics initiative noted that of 159 identified AI solutions, only a small subset had achieved regulatory clearance for clinical use in low- and middle-income countries; industry’s caution stems in part from insufficient independent evaluations.

Deployment logistics also require attention. A tool may perform well in controlled trials, yet front-line implementation in rural clinics demands robust usability, integration with workflows, data-privacy protections and health-worker training. Experts stress that AI must augment—not replace—clinical pathways and that over-reliance may introduce new risks of mis-triage, equity gaps or unintended bias.

Despite these caveats, the tools showcased at the conference reflect a shift in TB control strategy: from one-size diagnostics toward differentiated pathways, low-threshold community screening and data-driven targeting of interventions. The breath-based and cough-monitoring innovations carry particular promise for under-served regions where conventional lab and imaging infrastructure is scarce. At the same time the mapping tool points to smarter resource allocation rather than blanket screening efforts.

The success of these technologies will depend on real-world validation, alignment with national TB programmes, sustainable financing and ethical guardrails. If those alignments emerge, the integration of AI into the TB-care cascade could help bend the trajectory of a disease that remains the leading global infectious-killer.